# AeroZoo v1.2: 3,816 3D Incompressible RANS Flow Problems Across Seven Geometry Families (NVIDIA PhysicsNeMo)

## Facts

- Provider: Peter Sharpe, Carmelo Gonzales, Corey Adams, Mehdi Ataei, Sanjay Choudhry (NVIDIA)
- License: CC-BY-4.0 (https://huggingface.co/datasets/nvidia/AeroZoo)
- Size: 3,836 files
- Format: PNG, CSV (in sampled files)
- Stored at: Hugging Face, https://huggingface.co/datasets/nvidia/AeroZoo
- Access: public, readable without an account
- Category: Physics
- Source: https://arxiv.org/abs/2611.17152
- Page: https://scigantic.com/archive/3a1bca00-651b-51aa-ab82-5b0b1dec0fee
- Keywords: Physics, CFD, PNG, CSV, Hugging Face

## About

AeroZoo is a 3D aerodynamics dataset from NVIDIA (Sharpe, Gonzales, Adams, Ataei, Choudhry; arXiv:2611.17152) built to test whether neural surrogates solve boundary-value problems rather than memorising one shape family in one wind tunnel. It poses 3,816 distinct incompressible flow problems at Reynolds numbers from 1e3 to 1e8, solved with unsteady RANS using the Langtry-Menter transition SST model in OpenFOAM 14 and time-averaged over a fixed convective window. Bodies come from seven families (random blobs, Blender bluff and streamlined shapes, procedural cars, Zero-to-CAD parts, Airframe Foundry aircraft, streamlined bodies and wings) placed in eight enclosure configurations: rectangular boxes with open, ground-effect, confined, crossflow and dual-inlet layouts, opposed inlet/outlet spheres, symmetry hemispheres and 303 internal ducts, with no-slip or banded slip walls. Each sample stores the posed problem and time-mean solution as a PhysicsNeMo DomainMesh (memory-mapped TensorDict): about 23.8 billion interior query points in total (median 5.0M cells), surface pressure and wall shear stress, force coefficients, transition and stability labels (2,817 mean_stable), with 22 frozen train/validation/test splits for interior and boundary tasks. About 1.48 TB in tar shards, with index.json and SHA256SUMS. CC-BY-4.0, with 500 Zero-to-CAD parts under Apache-2.0; cite Sharpe et al. 2026.

## Load the data

```python
import pandas as pd
df = pd.read_csv("hf://datasets/nvidia/AeroZoo/zero-to-cad-provenance.csv")
df.head()
```

## Data structure

File types in sampled files: PNG, CSV

Example file: `zero-to-cad-provenance.csv`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/3a1bca00-651b-51aa-ab82-5b0b1dec0fee
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/3a1bca00-651b-51aa-ab82-5b0b1dec0fee
- Over MCP: `get_schema_card` with id `3a1bca00-651b-51aa-ab82-5b0b1dec0fee` at https://api.scigantic.com/mcp

## Related topics

- [CSV datasets](https://scigantic.com/topics/csv)
- [PNG datasets](https://scigantic.com/topics/png)
- [Physics datasets](https://scigantic.com/topics/physics)
- [CFD datasets](https://scigantic.com/topics/cfd)
